In the field of pavement monitoring, Ground Penetrating Radar (GPR) methods are gaining prominence due to their ability to perform non-destructive testing of the subsurface. In this context, the detection and characterization of subsurface debondings at an early stage is recommended to avoid further degradation and to maintain the lifespan of these structures. To mitigate the limited time resolution of the conventional GPR devices, this paper develops the detection of thin debonding (of millimeter-order) by monitoring small changes in the time domain GPR data by specific data processing techniques (with certain automatic capabilities). In this paper, we propose to use the supervised machine learning method, namely Two-class Support Vector Machines (SVM) to achieve the objectives. In addition, by means of time domain GPR signal features, we aim at reducing the computational burden and also increase the efficiency of SVM. The method is implemented to process independent 1D GPR A-scan data. Furthermore, the performance assessment of Two-class SVM is carried out on both simulated and field data by means of Sensitivity Analysis to identify the parameters that affect its performance. While simulated data is generated using the analytic Fresnel data model, the field data are UWB Stepped-Frequency GPR (SF-GPR) data which were collected over artificially embedded debondings. The data was acquired during the Accelerated Pavement Tests (APTs) conducted at the IFSTTAR's fatigue carousel to survey debonding growth in the defect-affected zones at various stages of fatigue. Two-class SVM presented the ability to detect thin millimetric debondings. Whereas, sensitivity analysis demonstrated a quick and efficient way to assess the pavement conditions.